See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3335914

Interval type-2 fuzzy logic systems: Theory and design

Article  in  IEEE Transactions on Fuzzy Systems · November 2000

DOI: 10.1109/91.873577 · Source: IEEE Xplore

CITATIONS READS
1,859 3,891

2 authors, including:

Jerry M Mendel
University of Southern California
547 PUBLICATIONS   48,327 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Comparing the Performance Potentials of Singleton and Non-Singleton Type-1 and Interval Type-2 Fuzzy Systems in Terms of Sculpting the State Space View project

Spatio-termporal characterization of responses of visual cells to natural images View project

All content following this page was uploaded by Jerry M Mendel on 22 August 2017.

The user has requested enhancement of the downloaded file.

IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000
Interval Type-2 Fuzzy Logic Systems:
Theory and Design
Qilian Liang and Jerry M. Mendel, Fellow, IEEE
Abstract—In this paper, we present the theory and design of
interval type-2 fuzzy logic systems (FLSs). We propose an efficient
and simplified method to compute the input and antecedent
operations for interval type-2 FLSs; one that is based on a general
inference formula for them. We introduce the concept of upper
and lower membership functions (MFs) and illustrate our efficient
inference method for the case of Gaussian primary MFs. We also
propose a method for designing an interval type-2 FLS in which
we tune its parameters. Finally, we design type-2 FLSs to perform
time-series forecasting when a nonstationary time-series is cor-
rupted by additive noise where SNR is uncertain and demonstrate
improved performance over type-1 FLSs.
Index Terms—Interval type-2 fuzzy sets, nonsingleton fuzzy logic
systems, time-series forecasting, tuning of parameters, type-2 fuzzy
logic systems, upper and lower membership functions.
I. INTRODUCTION
A FUZZY logic system (FLS) (also known as a fuzzy
system, fuzzy logic controller, etc) includes fuzzifier,
rules, inference engine, and defuzzifier [22]. Quite often,
the knowledge that is used to construct the rules in a FLS is
uncertain. Three ways in which such rule uncertainty can occur
are: 1) the words that are used in antecedents and consequents
of rules can mean different things to different people [23];
2) consequents obtained by polling a group of experts will
often be different for the same rule because the experts will
not necessarily be in agreement; and 3) noisy training data.
Antecedent or consequent uncertainties translate into uncertain
antecedent or consequent membership functions. Type-1 FLSs,
whose membership functions are type-1 fuzzy sets, are unable
to directly handle rule uncertainties. Type-2 FLSs, the subject
of this paper, in which antecedent or consequent membership
functions are type-2 fuzzy sets, can handle rule uncertainties.
The concept of type-2 fuzzy sets was introduced by Zadeh
[41] as an extension of the concept of an ordinary fuzzy set,
i.e., a type-1 fuzzy set. Type-2 fuzzy sets have grades of mem-
bership that are themselves fuzzy [4]. A type-2 membership
grade can be any subset in —the primary membership;
and, corresponding to each primary membership, there is a sec-
ondary membership (which can also be in ) that defines
the possibilities for the primary membership. A type-1 fuzzy
set is a special case of a type-2 fuzzy set; its secondary member-
ship function is a subset with only one element—unity. Type-2
Manuscript received August 20, 1999; revised May 30, 2000.
The authors are with the Signal and Image Processing Institute, Department
of Electrical Engineering Systems, University of Southern California, Los An-
geles, CA 90089-2564 USA (e-mail: mendel@sipi.usc.edu).
Publisher Item Identifier S 1063-6706(00)08455-1.
1063–6706/00$10.00 © 2000 IEEE
535
fuzzy sets allow us to handle linguistic uncertainties, as typi-
fied by the adage “words can mean different things to different
people.” A fuzzy relation of higher type (e.g., type-2) has been
regarded as one way to increase the fuzziness of a relation and,
according to Hisdal, “increased fuzziness in a description means
increased ability to handle inexact information in a logically
correct manner [6].”
Mizumoto and Tanaka [25] studied the set theoretic opera-
tions of type-2 sets, properties of membership grades of such
sets, and examined the operations of algebraic product and al-
gebraic sum for them [26]. More details about algebraic struc-
ture of type-2 sets are given in [28]. Dubois and Prade [2]–[4]
discussed fuzzy valued logic and give a formula for the compo-
sition of type-2 relations as an extension of the type-1 sup-star
composition, but this formula is only for minimum -norm. A
general formula for the extended sup-star composition of type-2
relations is given by Karnik and Mendel [11], [12], [17]. Based
on this formula, Karnik and Mendel [10]–[12], [17] established
a complete type-2 FLS theory to handle uncertainties in FLS pa-
rameters.
Similar to a type-1 FLS, a type-2 FLS includes fuzzifier, rule
base, fuzzy inference engine, and output processor. The output
processor includes type-reducer and defuzzifier; it generates a
type-1 fuzzy set output (from the type-reducer) or a crisp number
(from the defuzzifier). A type-2 FLS is again characterized by
IF–THEN rules, but its antecedent or consequent sets are now
type-2. Type-2 FLSs can be used when the circumstances are too
uncertain to determine exact membership grades such as when
training data is corrupted by noise—a case studied later in this
paper.
General type-2 FLSs are computationally intensive because
type-reduction is very intensive. Things simplify a lot when sec-
ondary membership functions (MFs) are interval sets (in this
case, the secondary memberships are either zero or one and we
call them interval type-2 sets) and this is the case studied in
this paper. When the secondary MFs are interval sets, we call
the type-2 FLSs “interval type-2 FLSs.” For some other discus-
sions on the use of interval sets in fuzzy logic (see Hisdal [6],
Schwartz [32], and Turksen [34]).
The most commonly used fuzzifier is a singleton; but, such
a fuzzifier is not adequate when data is corrupted by measure-
ment noise. In this case, a nonsingleton fuzzifier that treats each
measurement as a fuzzy number should be used. The theory and
applications of a type-1 FLS with nonsingleton fuzzifier are pre-
sented in [27], where the input is fuzzified into a type-1 fuzzy
set (e.g., Gaussian) whose parameters are based on the measured
input and the mean and variance of the measurement noise. This
assumes that the statistical knowledge (mean and variance) of
536 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

the noise is given or can be estimated; but, in many cases, these membership function of a fired rule can be expressed by
values are not known ahead of time and can not be estimated the following extended sup-star composition [12], [17]:
from the data. Instead, we only have some linguistic knowledge
about the noise, such as very noisy, moderately noisy, or approx- (1)
imately no noise. In this case, we cannot fuzzify the crisp input
as a type-1 fuzzy set, because type-1 MFs cannot fully repre- where is a -dimensional Cartesian product space,
sent the uncertainty associated with this linguistic knowledge. , is the measurement domain of input ,(
We believe that in this important case, the input should be fuzzi- ); and is given by
fied into a type-2 fuzzy set for use in a nonsingleton type-2 FLS.
This case is also studied in this paper. (2)
In this paper, we also provide a design method for an interval
type-2 FLS, where IF–THEN fuzzy rules are obtained from given Additionally
input–output (I/O) data. Two primary design tasks are structure
identification and parameter adjustment [7]. The former deter- (3)
mines input/output (I/O) space partition, antecedent and conse-
quent variables, the number of IF–THEN rules (which are deter- Substituting (3) and (2) into (1), the latter becomes
mined by the I/O space partition), and the number and initial loca-
tions of membership functions. The latter identifies a feasible set
of parameters under the given structure. In this paper, we focus
on parameter adjustments in an interval type-2 FLS. (4)
Tuning the parameters of a type-1 FLS is possible because
its output can be expressed as a closed-form mathematical Let
formula. Optimization methods for doing this have been exten-
sively studied (for example, [7], [19], [22], [24], and [36]). Un- (5)
fortunately, the output of a type-2 FLS cannot be represented
by a closed-form mathematical formula; hence, there is an addi- then
tional level of complexity associated with tuning its parameters.
To date, type-2 sets and FLSs have been used in decision
making [1], [40], solving fuzzy relation equations [35], survey (6)
processing, [12], [13], time-series forecasting [12], [14], func- Suppose that any of the rules in the FLS fire, where
tion approximation [12], time-varying channel equalization ; then, the output fuzzy set, for a type-2 FLS is
[17], control of mobile robots [39], and preprocessing of data [9].
In the sequel, results for a general type-2 nonsingleton fuzzy (7)
logic system (NSFLS) are given in Section II; meet and join op-
erations for interval sets are given in Section III; upper and lower For later use, we define
membership functions that characterize a type-2 MF are intro-
duced in Section IV; an efficient and simplified method to com- (8)
pute the input and antecedent operations for interval type-2 FLSs and
is given in Section V; type-reduction and defuzzification for an (9)
interval type-2 FLS are reviewed in Section VI; a method for de-
signing an interval type-2 FLS is given in Section VII; an applica- so that (6) can be re-expressed as
tion of our design method is given in Section VIII for time-series
forecasting of a nonstationary time-series that is corrupted by ad- (10)
ditive noise whose SNR is uncertain; and, finally, the conclusions
and topics for future research are given in Section IX. General type-2 FLSs are computationally intensive. Things
In this paper, denotes a type-1 fuzzy set; denotes simplify a lot when secondary MFs are interval sets, in which
the membership grade of in the type-1 fuzzy set ; de- case secondary memberships are either zero or one and, as we
notes a type-2 fuzzy set; denotes the membership grade demonstrate below, such simplifications make the use of type-2
of in the type-2 fuzzy set , i.e., , FLSs practical.
; denotes meet operation; and, denotes join
operation. Meet and join are defined and explained in great de- III. MEET AND JOIN FOR INTERVAL SETS
tail in [10]–[12], [15].
The meet and join operations, which are needed in (5)–(10),
can be greatly simplified for interval type-1 sets.
II. TYPE-2 FLSS: GENERAL RESULTS Theorem 1 (Meet of Interval Sets Under Minimum or Product
In a type-2 FLS with a rule base of rules in which each -Norms):
rule has antecedents, let the th rule be denoted by , where a) Let and be two interval
: IF is , is , , and is , THEN is . The type-1 sets (often called interval sets) with domains
LIANG AND MENDEL: INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS 537

( ), and (
), respectively. The meet between and ,
( ), under minimum or product
-norms (i.e., ) is given by

(11)

where .
b) The meet under minimum or product -norms of in-
terval type-1 sets , , having domains
, , , respectively, where
is an interval set with domain
, .
The proof of Theorem 1a), based on minimum or product op-
erations between two interval sets, is given in [18] and [32]. The
extension to part b) (via mathematical induction) is so straight-
forward, we leave it to the reader.
Theorem 2 (Join of Interval Sets):
a) Let and be as defined in part (a) of Theorem 1. The join
between and , ( ), is given by

(12)

where .
b) Let ( ) be as defined in Theorem 1(b).
Then the join of these interval type-1 sets is an interval
set with domain , .
The proof of Theorem 2(a), based on maximum operation
between two interval sets, is given in [18], [12], and [32].
The extension to part (b) (via mathematical induction) is
also so straightforward we leave it to the reader.
In this paper, we always assume that the operation is the
maximum operation.
Observe from Theorems 1 and 2, that meet and join opera-
tions of interval sets are determined just by the two end-points Fig. 1. The type-2 MFs for (a) Example 1 and (b) Example 2. The thick solid
of each interval set. In a type-2 FLS, the two end-points are asso- lines denote upper MFs and the thick dashed lines denote lower MFs. The shaded
ciated with two type-1 MFs that we refer to as upper and lower regions are the footprints of uncertainty for interval secondaries. In (a), the
centers of Gaussian MFs vary from 4.5–5.5; in (b), the center of the Gaussian
MFs. MFs is 5 and the variance varies from 1.0–2.0.

IV. UPPER AND LOWER MFS FOR TYPE-2 FLSS

We use an overbar (underbar) to denote the upper (lower) MF.
For convenience in defining the upper and lower MFs of a For example, the upper and lower MFs of are
type-2 MF, we first give the definition of footprint of uncertainty and , respectively, so that
of a type-2 MF.
Definition 1 (Footprint of Uncertainty of a Type-2
MF): Uncertainty in the primary membership grades of a (13)
type-2 MF consists of a bounded region that we call the
footprint of uncertainty of a type-2 MF (e.g., see Fig. 1). It is
the union of all primary membership grades. Similarly, we will represent and as
Definition 2 (Upper and Lower MFs): An upper MF and a
lower MF are two type-1 MFs that are bounds for the footprint of
(14)
uncertainty of an interval type-2 MF. The upper MF is a subset
that has the maximum membership grade of the footprint of un-
certainty; and the lower MF is a subset that has the minimum (15)
membership grade of the footprint of uncertainty.
538 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

Example 1: Gaussian Primary MF with Uncertain Example 2: Gaussian Primary MF with Uncertain Standard
Mean: Consider the case of a Gaussian primary MF having a Deviation: Consider the case of a Gaussian primary MF having
fixed standard deviation and an uncertain mean that takes a fixed mean and an uncertain standard deviation that takes
on values in , i.e., on values in , i.e.,

(16) (19)
where where
; ;
number of antecedents; number of antecedents;
; and, ;
number of rules. number of rules.
The upper MF is [see Fig. 1(a)] The upper MF is [see Fig. 1(b)]
(20)
(17) and the lower MF is [see Fig. 1(b)]
(21)
where, for example Note that the upper and lower membership functions are sim-
pler for Example 2 than for Example 1.
These examples illustrate how to define and so it is clear
how to define these membership functions for other situations
The lower MF is [see Fig. 1(a)] (e.g., triangular, trapezoidal, bell MFs).

V. INTERVAL TYPE-2 FLSS

(18) Our major result for interval type-2 FLSs is given in:
Theorem 3: In an interval type-2 nonsingleton FLS with
type-2 fuzzification and meet under minimum or product
-norm: 1) the result of the input and antecedent operations,
in (9), is an interval type-1 set, i.e., , where
From this example, we see that sometimes an upper (or a
lower) MF cannot be represented by one mathematical function
over its entire domain. It may consist of several branches each
defined over a different segment of the entire domain. When
(22)
the input is located in one domain-segment, we call its cor-
responding MF branch an active branch, e.g., in Example 1, and
when , the active branch for
is .
When an upper (or lower) MF is represented in different seg- (23)
ments, its left-hand and right-hand derivatives at the segment
end point [e.g., for ] may not the supremum is attained when each term in brackets attains its
be equal, so the upper (or lower) MF may not be differentiable supremum; 2) the rule fired output consequent set in
over the entire domain; however, it is piecewise differentiable, (10) is
i.e., each branch is differentiable over its segment domain. This
fact will be used by us when we tune the parameters of a type-2 (24)
FLS.
Some upper and lower MFs can be represented by one func- where and are the lower and upper membership
tion and are differentiable over their entire domain as we demon- grades of ; and 3) the output fuzzy set in (7) is
strate in the following example. (25), as shown at the bottom of the page.

(25)
LIANG AND MENDEL: INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS 539

A. Proof of Theorem 3 The suprema in (31) and (32) are, overall, in . By

1) Applying Theorem 1(a) to (5) for an interval type-2 FLS the monotonicity property of a -norm [42], [27], the
with type-2 fuzzifier and using (14) and (15), we find supremum is attained when each term in brackets attains
its supremum.
2) Based on (9), (31), (32), and Theorem 2(a), we evaluate
(10) as

(26)
where . So, the meet between an input type-2
set and an antecedent type-2 set just involves the -norm (33)
operation between the points in two upper or lower MFs.
The upper and lower MFs of are
3) Because ( ) are interval sets, it is
straightforward to obtain in (7) using Theorem
(27)
2(b). The result is (25).

(28)
In evaluating (22) and (23), the supremum is attained when
each term in brackets attains its supremum; so, in the inference
The meet operations in (8) are in a -dimensional of a type-2 FLS, we will examine
Cartesian product space so the meet operation is over all
points , . Based on Theorem 1(b),
(34)
we know that the upper membership grades of ,
(a type-1 MF) are obtained from the -norm of
membership grades in ; hence, from (27), we find (35)

where , and is a -norm; then, and can be

re-expressed as
(29)
(36)
The lower membership grades (a type-1 MF), are (37)
the -norm of the membership grades in ; hence,
from (28) we find where denotes -norm. We illustrate (36) and (37) below in
Section V-C.

(30)
The join operation in (9) is over all points in . Based
on Theorem 2, we know that the right-most point of the
join of ( ) interval sets is the maximum value
of all the right-most points in the interval sets; so, the
right-most point of comes from the maximum value
(supremum) of (the right-most point of interval set
for each value of ); hence, from (29) we find
B. Corollaries to Theorem 3
When the input is fuzzified to a type-1 fuzzy set so that
( ), the upper and lower MFs of
merge into one MF in which case Theorem 3
simplifies to the following.
Corollary 1: In an interval type-2 FLS with nonsingleton
type-1 fuzzification and meet under minimum or product
-norm, and in (22) and (23) simplify to

(38)

(31) and

Similarly, the left-most point of comes from the

maximum value (supremum) of ; hence, from (30) (39)
we find
where ( ) is the type-1 fuzzified input.
When a singleton fuzzifier is used, the upper and lower MFs
of merge into one crisp value, namely one, in which
(32) case Theorem 3 simplifies further to the following.
540 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

Fig. 2. Type-2 FLS: input and antecedent operations. (a) Singleton fuzzification with minimum t-norm; (b) singleton fuzzification with product t-norm; (c) NS
type-1 fuzzification with minimum t-norm; (d) NS type-1 fuzzification with product t-norm; (e) NS type-2 fuzzification with minimum t-norm; and (f) NS type-2
fuzzification with product t-norm. The dark shaded regions depict the meet between input and antecedent [computed using Theorem 1(a)].

Corollary 2: In an interval type-2 FLS with singleton fuzzi- C. Illustrative Examples

fication and meet under minimum or product -norm and
Example 3—Pictorial Representation of Input and An-
in (22) and (23) simplify to
tecedent Operations: In Fig. 2, we plot the results of input and
(40) antecedent operations with singleton, type-1 nonsingleton, and
type-2 nonsingleton fuzzifications. The number of antecedents
and is . In all cases, the firing stength is an interval type-1
(41) set, , where and . For
singleton fuzzification [Fig. 2(a) and (b)], denotes the
where ( ) denotes the location of the singleton.
firing strength between input and , namely ;
The proofs of these corollaries are so simple, we leave them
for the reader. and denotes the firing strength between input and ,
LIANG AND MENDEL: INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS 541

TABLE I
x FOR EXAMPLE 4

TABLE II
x FOR EXAMPLE 4 BASED ON PRODUCT t-NORM

namely , , as established by Corollary by , and by . The th antecedent MF has the fol-

2. For nonsingleton type-1 fuzzification [Fig. 2(c) and (d)], lowing form:
denotes the supremum of the firing strength between the
-norm of membership functions and ; and denotes
the supremum of the firing strength between the -norm of
membership functions and , , as established
(43)
by Corollary 1. For nonsingleton type-2 fuzzification [Fig. 2(e)
and (f)], denotes the supremum of the firing strength and its upper and lower MFs and are obtained
between the -norm of upper membership functions and
; and, denotes the supremum of the firing strength from (17) and (18), respectively, by replacing by ,
between the -norm of lower membership functions and by and by . Observe that there are six pa-
, , as established by Theorem 3. The main thing rameters that determine these two type-2 Gaussian MFs: ,
, , , , and . In this example, as in [27],
to observe from these figures is that regardless of singleton or
nonsingleton fuzzification and minimum or product -norm, the we assume that
result of input and antecedent operations is an interval type-1
(44)
set that is determined by its left-most point and right-most
point . and our objective is to evaluate (34) and (35). Equation (44)
Example 4—Input is a Gaussian Primary MF with Uncer- means that uncertainty in each input set is always no more than
tain Standard Deviation and Antecedents are Gaussian Primary the uncertainty in the antecedent.
MFs with Uncertain Means: In this example, we compute We denote the value of at which the supremum of (34)
and when a Gaussian primary MF with an uncertain stan- occurs as and the value of at which the supremum of
dard deviation (as in Example 2) is used as input fuzzy sets and (35) occurs as . The results for and of this
Gaussian primary MFs with uncertain means (as in Example example are carried out in Appendix A, and are summarized in
1) are used as antecedent MFs. This case is important to our Tables I–III. From these results, it is straightforward to compute
time-series forecasting application in Section VIII. In this case and using (34) and (35), i.e.,

(45)
(46)
(42)
When the input is fuzzified to a type-1 Gaussian MF, then
and its upper and lower MFs and are obtained , and we can easily obtain and
from (20) and (21), respectively, by replacing by , based on Tables I–III. When a singleton fuzzifier is used, the
542 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

TABLE III
x FOR EXAMPLE 4 BASED ON MINIMUM t-NORM

results in Tables I–III simplify even further since centroid of the type-2 interval conse-
. and quent set (the centroid of a type-2
fuzzy set is described in [11], [16], and
VI. TYPE REDUCTION AND DEFUZZIFICATION [12]);
.
A type-2 FLS is a mapping . After fuzzification, Observe, that each set on the right-hand side (RHS) of (47) is an
fuzzy inference, type-reduction, and defuzzification, we obtain interval type-1 set, hence,
a crisp output. For an interval type-2 FLS, this crisp output is the is also an interval type-1 set. So, to find
center of the type-reduced set. Based on Theorem 3 and Corol- , we just need to compute the two
laries 1 and 2, we know that for an interval type-2 FLS, regard- end-points of this interval. Unfortunately, no closed-form
less of singleton or nonsingleton fuzzification and minimum or formula is available for .
product -norm, the result of input and antecedent operations For any value , can be represented as
(firing strength) is an interval type-1 set, which is determined by
its left-most and right-most points and (e.g., see Fig. 2).
The fired output consequent set of rule can be ob-
tained from the fired interval strength using (24) or Corollaries
(48)
1 or 2 and (24). Then the fired combined output consequent set
can be computed using (25).
Type-reduction was proposed by Karnik and Mendel [11],
[12], [17]. It is an “extended version” [using the extension prin-
the maximum value of is and the minimum value of is
ciple [41] of type-1 defuzzification methods and is called type-
. From (48), we see that is a monotonic increasing function
reduction because this operation takes us from the type-2 output
with respect to ; so is associated only with and, similarly,
sets of the FLS to a type-1 set that is called the “type-reduced
is associated only with . In the center of sets (COS)-type-
set.” This set may then be defuzzified to obtain a single crisp
reduction method, Karnik and Mendel [12], [17] have shown
number; however, in many applications, the type reduced set
that the two end points of , , and depend only on a
may be more important than a single crisp number since it con-
veys a measure of uncertainties that have flown through the mixture of or values, since . In this
type-2 FLS. case, and can each be represented as a fuzzy basis function
There exist many kinds of type-reduction, such as centroid, (FBF) expansion, i.e.,
center-of-sets, height, and modified height, the details of which
are given in [11], [12], and [17]. In this paper, for illustrative
purposes, we focus on center-of-sets type-reduction, which can
be expressed as (49)

where denotes the firing strength membership grade

(47) [either or ] contributing to the left-most point and
is the FBF. Similarly

where
interval set determined by two end (50)
points and ;
;
LIANG AND MENDEL: INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS 543

where denotes the firing strength membership grade (either VII. DESIGNING INTERVAL TYPE-2 FLSS BASED ON TUNING
or ) contributing to the right-most point and
Given an input–output training pair , and
is another FBF. , we wish to design an interval type-2 FLS with output (53) so
Note that whereas a type-1 FLS is characterized by a single that the error function
FBF expansion [22], [37], an interval type-2 FLS is character-
ized by two FBF expansions. A general type-2 FLS is charac- (54)
terized by a huge number of FBF expansions [12], [17]; hence,
we have demonstrated that by choosing secondary membership is minimized. Based on the analysis in Section VI, we know that
functions to be interval sets, the complexity of a general type-2 only the upper and lower MFs and the two endpoints of (the
FLS is vastly reduced. center of the consequent set) determine . So we want to
In order to compute and , we need to compute tune the upper and lower MFs and the consequent parameters
and . This can be done . Since an interval type-2 FLS can be character-
using the exact computational procedure given in [12], [16], and ized by two FBF expansions that generate the points and ,
[17]. Here, we briefly provide the computation procedure for respectively, we can focus on tuning the parameters of just these
two type-1 FLSs.
. Without loss of generality, assume the s are arranged in
Given input–output training samples ,
ascending order, i.e., .
( ), we wish to update the design parameters
1) Compute in (50) by initially setting so that (54) is minimized for training epochs (updating the
for , where and have been previously parameters using all the training samples one time is called
computed using (22) and (23) and let . “one epoch”). A general method for doing this is as follows.
2) Find ( ) such that . 1) Initialize all the parameters including the parameters in
3) Compute in (50) with for and antecedent and consequent MFs and input sets.
for and let . 2) Set the counter of training epoch .
4) If , then go to Step 5). If , then stop and 3) Set the counter of training data sample .
4) Apply input to the type-2 FLS, and compute the
set .
total firing degree for each rule, i.e., compute and
5) Set equal to and return to Step 2).
( ) using Theorem 3.
This four-step computation procedure [Step 1) is an initial- 5) Compute and , as described in Section VI (which
ization step] has been proven to converge to the exact solution leads to a reordering of the rules; but, they are then
in no more than iterations [12]. Observe that in this proce- renumbered 1, 2, , ). This will establish and ,
dure, the number is very important. For , , and so that and can be expressed as
for , ; so can be represented as

(51)

The procedure for computing is very similar. Just replace (55)

by and, in Step 2), find ( ) such that
and, in step 3, let for , and
for . Then can be represented as

(52)
(56)
Because is an interval set, we defuzzify it using the av-
erage of and ; hence, the defuzzified output of an interval
type-2 FLS is
6) Compute , which is the defuzzified
output of the type-2 FLS.
(53) 7) Determine the explicit dependence of and on
membership functions (because and obtained in
A perfect FLS should have , where is the de- Step 5) may have changed from one iteration to the
sired output but, generally, there exist errors between the desired next, the dependence of and on MFs may also have
output and actual output. We, therefore, need a design procedure changed). To do this, first determine the dependence of
for tuning the parameters of the FLS in order to minimize such and on membership functions, using (34)–(37),
errors. i.e., is determined by , , ,
544 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

, , and, is determined by VIII. FORECASTING OF CHAOTIC TIME SERIES

, , , , . Type-1 FLSs have been extensively used in time-series fore-
Consequently casting (e.g., [7], [8], [22], [24]). Here, we evaluate the perfor-
mance of our design method by applying it to the forecasting
of a Mackey–Glass chaotic time-series. We compare the perfor-
mance of interval type-2 designs with that of type-1 designs.

A. Mackey–Glass Chaotic Time Series

The Mackey–Glass chaotic time series can be represented as
[20]

(59)

When , (59) exhibits chaotic behavior. In simulating

(59), we converted it to a discrete-time equation by using Euler’s
method [30]. Denoting
(57)
(60)
Similarly
then

(61)

where is a small number and the initial values of

are set randomly. We chose and .
(58) In our simulations, we assumed that the sampled time-se-
ries is corrupted by uniformly distributed additive noise
and only noisy measured values of are available, i.e.,
8) Test each component of to determine the active
, . In actual time-series
branches in , , and ,
such as the price curve for the U.S. dollar versus the German
, and represent the active branches mark, market volatility can change noticeably over the course
as functions of their associated parameters, e.g., use of time, so that the variance of the noise component, which is
Tables I–III for the membership functions described in
related to volatility, need not be constant [21]. In our simula-
Example 4. This step depends on the locations of in
tion, we therefore assumed the noise is zero mean but has a
relation to the MFs.
signal-to-noise ratio (SNR) ranging from 0 dB [with standard
9) Tune the parameters of the active branches and the pa-
deviation (std) )] to 10 dB (with std ); it is uni-
rameters in the consequent using a steepest-descent (or
formly distributed into 100 levels from to and
any other optimization) method. The error function is
(54).
10) Set . If , go to Step 11); otherwise go to (62)
Step 4).
11) Set . If , stop; otherwise, go to Step 3). Our simulations were based on points, ,
, , . The first 500 data , ,
When the input is fuzzified to a type-1 fuzzy set, (57) and (58)
, are for training and the remaining 500 data
are simplified since , or,
, are for testing. In Fig. 3(a),
if a singleton fuzzifier is used, (57) and (58) are simplified even
further since all and disappear in (57) and we plot the Mackey–Glass chaotic time series ,
, , and, in Fig. 3(b), we plot one realization
(58), because and are two singletons.
of the noise corrupted data , , , .
What makes the tuning of the parameters of this type-2 FLS
challenging and different from the tuning of the parameters in a
B. Simulations
type-1 FLS is having to first determine which parameters and
depend on. This requires comparing ( ) to We used four antecedents for forecasting, i.e., ,
some points associated with parameters of the upper and lower , , and were used to predict . As in [7],
antecedent membership functions (e.g., as in Example 4). As we used only two fuzzy sets for each antecedent, so the number
these parameters change due to their tuning, it is highly likely of rules is . The initial locations of antecedent MFs were
that the dependency of and on parameters also changes. based on the mean and std of the first 500 points ,
This behavior does not occur in a type-1 FLS. , , .
LIANG AND MENDEL: INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS
Fig. 3. The Mackey–Glass chaotic time series shows (a) the noise-free data s(1001), s(1002), . . ., s(2000) and (b) the noise corrupted data (in one realization),
x(1001), x(1002), . . ., x(2000).
We compared the performance of the following five
forecasting FLSs: type-1 singleton FLS (SFLS); type-1 non-
singleton FLS (NSFLS),; type-2 SFLS; type-2 NSFLS with
type-1 fuzzifier (type-2 NSFLS-1); and type-2 NSFLS with
type-2 fuzzifier (type-2 NSFLS-2). Gaussian MFs were chosen
for the antecedents of the type-1 FLSs; the Gaussian MFs
of Example 1 were chosen for the antecedents of the type-2
FLSs; type-1 Gaussian MFs were chosen for the inputs of
the type-1 NSFLS and type-2 NSFLS with type-1 fuzzifier
and the Gaussian MFs of Example 2 were chosen for the
inputs of the type-2 NSFLS with type-2 fuzzifier. The pa-
545
rameters and number of parameters in these five FLSs are
summarized in Table IV [e.g., for the type-2 SFLS, total # of
P ]. The initial values of all the design
parameters are summarized in Table V.
After training, the rules were fixed and we tested the fuzzy
logic (FL) forecaster based on the remaining 500 noisy points,
, , , .
For each of the five designs, we ran 50 Monte-Carlo real-
izations and, for each realization, each FLS was tuned using
a simple steepest-descent algorithm for six epochs. After each
epoch, we used the testing data to see how each FLS performed
546 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

TABLE IV
THE PARAMETERS AND NUMBER OF PARAMETERS IN FIVE DIFFERENT FLSS (WITH 16 RULES AND FOUR ANTECEDENTS IN EACH RULE,
I.E., i = 1; . . . ; 16, AND k = 1; . . . ; 4). P IS SHORT FOR PARAMETERS

TABLE V
INITIAL VALUES OF THE PARAMETERS IN FIVE FLSS. m = m = x FOR ALL NSFLS DESIGNS. EACH ANTECEDENT IS DESCRIBED BY TWO FUZZY SETS

by evaluating the rmse between the defuzzified output of each in adaptive filters such as channel equalizers (e.g., [29],
FLS and the noise-free data, i.e., [31], [38]) because such equalizers must be robust to ad-
ditive noise.

rmse (63) IX. CONCLUSIONS AND FUTURE WORK

where , and
denotes transpose.
Since there are rmse values for each design,
we summarize the mean and std of the rmses for each epoch and
for each design in Fig. 4(a) and (b). Observe the following from
the figures.
1) Type-2 FLSs outperform type-1 FLSs. The type-2
NSFLS-2 performs the best and the type-2 NSFLS-1
also gives very good results. The reason for the latter
is because the type-2 NSFLS-1 uses in (62) as the
initial std of the input sets and this value for gives a
good approximation to the average value of the std of the
uniform noise.
2) The type-2 FLSs (especially type-2 NSFLS-2) achieve
close to their optimal performance almost at the first
epoch of tuning. This property shows that type-2 FLSs
(as compared to type-1 FLSs) are very promising for
real-time signal processing where more than one epoch
of tuning is not possible.
3) From the std of the rmses, we see that the type-2 FLSs (es-
pecially the type-2 NSFLS-2) have a considerably smaller
std than do type-1 FLSs, which demonstrates that type-2
FLSs are much more robust to the noise than are type-1
FLSs. Hence, type-2 FLSs appear to be promising for use
We have presented the theory and design of interval type-2
FLSs, including an efficient and simplified method to compute
their input and antecedent operations. We have also provided
a method for tuning the parameters of an interval type-2 FLS.
Our simulation results show that an interval type-2 FLS outper-
forms a type-1 FLS in forecasting a chaotic time-series whose
measurements were corrupted by nonstationary noise.
Interval type-2 FLSs provide a way to handle knowledge
uncertainty. Data mining and knowledge discovery are impor-
tant research topics that are being studied by researchers of
neural networks, fuzzy logic systems, evolutionary computing,
soft computing, artificial intelligence, etc. We believe that
interval type-2 FLSs have the potential to solve data mining and
knowledge discovery problems in the presence of uncertainty.
As we discussed, the large number of parameters and associ-
ated training complexity are the main disadvantages of interval
type-2 FLSs. The challenge is to develop ways to reduce the
number of parameters and training complexity. Some ways for
reducing the number of design parameters that remain to be ex-
plored are: 1) assume the subsets in each antecedent are the same
for all rules and 2) fix some parameters such as those in the
membership functions of measured inputs.
We believe that promising areas in which type-2 FLSs may
be advantageous over type-1 FLSs include mobile communi-
cations, communication networks, pattern recognition, and ro-
bust control because lots of uncertain information needs to be
LIANG AND MENDEL: INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS 547

Mouzouris and Mendel in [27] for the type-1 nonsingleton

case; they derive the value of at which
is achieved, where and
are Gaussians. Denoting this value of as ,
they have shown that for product -norm

(64)

and for minimum -norm

(65)

A. Determination of
To determine , we need to consider three cases of dif-
ferent locations:
1) : This case is depicted in Fig. 5(a) (the solid
lines). As we see from (17), the active branch in the upper
MF of the antecedent comes from

so it is very straightforward to obtain based on

(64) or (65). The result is given in Case 1 in Table I. It is
easy to verify that for both product and
minimum -norm.
2) , . This case is depicted in
Fig. 5(b)–(d) (these three possible cases are needed to
determine , as we explain below, but the same
result is obtained for in all three cases). As
we see from Example 1, in all three cases, the active
branch in the upper MF of the antecedent comes from the
constant MF and, since ,
Fig. 4. The mean and std of the rmses (for the test data) for the five FLS so , which leads to
designs averaged over 50 Monte-Carlo realizations. Tuning was performed in
each realization for six epochs. (a) Mean values. (b) STD values. . This result is given in Case 2 in
Table I.
3) : This case is depicted in Fig. 5(e). The
handled in these areas. For example, in mobile communica-
analysis for this case is very similar to that in Part I and
tions, the fading channel coefficients are uncertain during ad-
the result is given in Case 3 in Table I.
jacent training periods. Existing algorithms treat them as cer-
tain. In communication networks, Tanaka and Hosaka [5], [33]
observed the difficulties of obtaining appropriate MFs for effi- B. Determination of
cient network control, which suggests that type-2 MFs will be a 1) Product -Norm: To determine under product
better way to represent the uncertainty in such a network. -norm, we again need to consider three cases of locations.
Finally, whether or not design procedures can be developed As we see from (18), the lower MF of the antecedent comes
for noninterval type-2 FLSs remains to be explored. from the intersection of two Gaussian MFs and that intersection
point is . When , the
APPENDIX A active branch is , and, when
DETAILS OF EXAMPLE 4 , the active branch is , .
1) For the cases depicted in Fig. 5(a) and (b), in-
In determining and , we need the calculation
volves the lower MF and ( , ; ), i.e.,
of the supremum of a product or minimum of two type-1
Gaussian MFs. Such a calculation has been carried out by . For
548 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

Fig. 5. In Example 4, possible locations of input type-2 MF and antecedent type-2 MF. The MF with the large footprint and variance is the antecedent MF, and
the other is the input MF. (a)–(e) Cases 1-5, respectively.

this point to be located to the left of , This result is given in Case 1 in Table II.
we require 2) For the cases depicted in Fig. 5(d) and (e), in-
volves the lower MF and ( , ; ). The
(66) analysis for this part is very similar to that for item 1 and
we give the result in Table II, Case 3.
3) For the case depicted in Fig. 5(c) based on our re-
so that
sults in Table II, Cases 1 and 3,

(67)
. In this case, the
LIANG AND MENDEL: INTERVAL TYPE-2 FUZZY LOGIC SYSTEMS 549

supremum point of the product between the lower MF [18] A. Kaufmann and M. M. Gupta, Introduction to Fuzzy Arithmetic:
and ( , ; ) is located to the right of Theory and Applications. New York: Van Nostrand Reinhold, 1991.
[19] C.-T. Lin and C. S. G. Lee, Neural Fuzzy Systems. Englewood Cliffs,
or at (recall Case 1), but , NJ: Prentice-Hall, 1996.
is only active when . [20] M. C. Mackey and L. Glass, “Oscillation and chaos in physiological
control systems,” Sci., vol. 197, pp. 287–289, 1987.
As we know, the product of two Gaussian MFs is convex [21] M. Magdon-Ismail, A. Nicholson, and Y. Abu-Mostafa, “Financial
and, to the left of the supremum point, it is monotonically markets: Very noisy information processing,” Proc. IEEE, vol. 86, pp.
increasing. Hence, the supremum point of the product 2184–2195, Nov. 1998.
[22] J. M. Mendel, “Fuzzy logic systems for engineering: A tutorial,” Proc.
between the lower MF and ( , ; ) IEEE, vol. 83, pp. 345–377, Mar. 1995.
[ ] is at . [23] , “Computing with words when words can mean different things
to different people,” in Int. ICSC Congress Computat. Intell.: Methods
Additionally, the supremum point of the product be- Applicat., 3rd Annu. Symp. Fuzzy Logic Applicat., Rochester, NY, June
tween the lower MF and ( , , ), is lo- 1999.
cated to the left of or at (recall Case 2). [24] J. M. Mendel and G. Mouzouris, “Designing fuzzy logic systems,” IEEE
Trans. Circuits Syst. II, vol. 44, pp. 885–895, Nov. 1997.
Hence, similar to the above analysis, the supremum point [25] M. Mizumoto and K. Tanaka, “Some properties of fuzzy sets of type 2,”
of the product between the lower MF and ( , Inform. Contr., vol. 31, pp. 312–340, 1976.
[26] , “Fuzzy sets of type 2 under algebraic product and algebraic sum,”
)[ ] is at . Fuzzy Sets Syst., vol. 5, pp. 277–290, 1981.
Hence, [27] G. C. Mouzouris and J. M. Mendel, “Nonsingleton fuzzy logic systems:
Theory and application,” IEEE Trans. Fuzzy Syst., vol. 5, pp. 56–71, Feb.
1997.
(68) [28] J. Nieminen, “On the algebraic structure of fuzzy sets of type-2,” Kyber-
netica, vol. 13, no. 4, Feb. 1977.
This result is given in Table II, Case 2. [29] S. K. Patra and B. Mulgrew, “Efficient architecture for Bayesian equal-
ization using fuzzy filters,” IEEE Trans. Circuits Syst. II, vol. 45, pp.
2) Minimum -Norm: The results to obtain under 812–820, July 1998.
minimum -norm are given in Table III. The steps to obtain these [30] D. Quinney, An Introduction to the Numerical Solution of Differential
Equation. Hertforshire, U.K.: Research Studies, 1985.
results are exactly the same as those used to obtain the results [31] P. Sarwal and M. D. Srinath, “A fuzzy logic system for channel equal-
in Table II and are left to the reader. ization,” IEEE Trans. Fuzzy Syst., vol. 3, pp. 246–249, May 1995.
[32] D. G. Schwartz, “The case for an interval-based representation of lin-
guistic truth,” Fuzzy Sets Syst., vol. 17, pp. 153–165, 1985.
REFERENCES [33] Y. Tanaka and S. Hosaka, “Fuzzy control of telecommunications net-
[1] J. L. Chaneau, M. Gunaratne, and A. G. Altschaeffl, “An application works using learning technique,” Electron. Communicat. Japan, pt. I,
of type-2 sets to decision making in engineering,” in Analysis of Fuzzy vol. 76, no. 12, pp. 41–51, Dec. 1993.
Information—Vol. II: Artificial Intelligence and Decision Systems, J. C. [34] I. Turksen, “Interval valued fuzzy sets based on normal forms,” Fuzzy
Bezdek, Ed. Boca Raton, FL: CRC, 1987. Sets Syst., vol. 20, pp. 191–210, 1986.
[2] D. Dubois and H. Prade, “Operations on fuzzy numbers,” Int. J. Syst. [35] M. Wagenknecht and K. Hartmann, “Application of fuzzy sets of type 2
Sci., vol. 9, pp. 613–626, 1978. to the solution of fuzzy equation systems,” Fuzzy Sets Syst., vol. 25, pp.
[3] , “Operations in a fuzzy-valued logic,” Inform. Contr., vol. 43, pp. 183–190, 1988.
224–240, 1979. [36] L.-X. Wang and J. M. Mendel, “Back-propagation of fuzzy systems as
[4] , Fuzzy Sets and Systems: Theory and Applications. New York: nonlinear dynamic system identifiers,” in Proc. IEEE Int. Conf. Fuzzy
Academic, 1980. Syst., San Diego, CA, Mar. 1992, pp. 1409–1418.
[5] S. Ghosh, Q. Razouqi, H. J. Schumacher, and A. Celmins, “A survey of [37] , “Fuzzy basis functions, universal approximation, and orthogonal
recent advances in fuzzy logic in telecommunications networks and new least squares learning,” IEEE Trans. Neural Networks, vol. 3, pp.
challenges,” IEEE Trans. Fuzzy Syst., vol. 6, pp. 443–447, Aug. 1998. 807–814, Sept. 1992.
[6] E. Hisdal, “The IF THEN ELSE statement and interval-valued fuzzy [38] , “Fuzzy adaptive filters, with application to nonlinear channel
sets of higher type,” Int. J. Man–Machine Studies, vol. 15, pp. 385–455, equalization,” IEEE Trans. Fuzzy Syst., vol. 1, pp. 161–170, Aug. 1993.
1981. [39] K. C. Wu, “Fuzzy interval control of mobile robots,” Comput. Elect.
[7] J.-S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference Eng., vol. 22, no. 3, pp. 211–229, 1996.
system,” IEEE Trans. Syst., Man, Cybern., vol. 23, pp. 665–685, [40] R. R. Yager, “Fuzzy subsets of type II in decisions,” J. Cybern., vol. 10,
May/June 1993. pp. 137–159, 1980.
[8] J.-S. R. Jang and C. -T. Sun, “Neuro-fuzzy modeling and control,” Proc. [41] L. A. Zadeh, “The concept of a linguistic variable and its application to
IEEE, vol. 83, pp. 378–406, Mar. 1995. approximate reasoning—I,” Inform. Sci., vol. 8, pp. 199–249, 1975.
[9] R. I. John, P. R. Innocent, and M. R. Barnes, “Type 2 fuzzy sets and [42] H. J. Zimmermann, Fuzzy Set Theory and Its Applications, 2nd
neuro-fuzzy clustering of radiographic tibia images,” in Proc. 1998 ed. Boston, MA: Kluwer, 1991.
IEEE Int. Conf. Fuzzy Syst., Anchorage, AK, May 1998, pp. 1373–1376.
[10] N. N. Karnik and J. M. Mendel, “Introduction to type-2 fuzzy logic sys-
tems,” in Proc. IEEE FUZZ Conf., Anchorage, AK, May 1998.
[11] , “Type-2 fuzzy logic systems: Type-reduction,” in IEEE Syst.,
Man, Cybern. Conf., San Diego, CA, Oct. 1998.
[12] , “An introduction to type-2 fuzzy logic systems,”, USC Report, Qilian Liang received the B.S. degree from Wuhan
http://sipi.usc.edu/~mendel/report, Oct. 1998. University, China, in 1993, the M.S. degree from
[13] , “Applications of type-2 fuzzy logic systems: handling the un- Beijing University of Posts and Telecommunica-
certainty associated with surveys,” in Proc. FUZZ-IEEE Conf., Seoul, tions, China, in 1996, and the Ph.D. degree from
Korea, Aug. 1999. University of Southern California, Los Angeles, in
[14] , “Applications of type-2 fuzzy logic systems to forecasting of time- 2000, all in electrical engineering.
series,” Inform. Sci., vol. 120, pp. 89–111, 1999. He was a Research Assistant in the Department
[15] , “Operations on type-2 fuzzy sets,” Fuzzy Sets Syst., 2000, to be of Electrical Engineering-Systems, University of
published. Southern California, Los Angeles, from September
[16] N. N. Karnik, J. M. Mendel, and Q. Liang, “Centroid of a type-2 fuzzy 1997 to May 2000. His research interests include
set,” Inform. Sci., submitted for publication. wireless communications, broad-band networks,
[17] , “Type-2 fuzzy logic systems,” IEEE Trans. Fuzzy Syst., vol. 7, pp. fuzzy logic systems, and video traffic classification. He is currently with
643–658, Dec. 1999. Hughes Network Systems, San Diego, CA.
550 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 8, NO. 5, OCTOBER 2000

Jerry M. Mendel (S’59–M’61–SM’72–F’78) re-

ceived the Ph.D. degree in electrical engineering from
the Polytechnic Institute of Brooklyn, Brooklyn, NY.
Currently, he is Professor of Electrical Engi-
neering and Associate Director for Education of the
Integrated Media Systems Center at the University
of Southern California, Los Angeles, where he
has been since 1974. He has published over 380
technical papers and is author and/or editor of seven
books. His present research interests include type-2
fuzzy logic systems and their applications to a wide
range of problems.
Dr. Mendel is a Distinguished Member of the IEEE Control Systems Society.
He was President of the IEEE Control Systems Society in 1986. Among his
awards are the 1983 Best Transactions Paper Award of the IEEE Geoscience
and Remote Sensing Society, the 1992 Signal Processing Society Paper Award,
a 1984 IEEE Centennial Medal, and an IEEE Third Millenium Medal.

View publication stats